FEAST: Federated Activation Steering of Protein Language Models for Collaborative Protein Sequence Optimization

The drive for better performance in protein sequence optimization often motivates collaboration across organizations that share design goals, yet privacy and intellectual property constraints often prohibit collaborative sharing of proprietary data. While activation steering offers a lightweight method to guide frozen Protein Language Models (PLMs) using local data, generalizing this to such a collaborative setting is challenging: directly aggregating steering directions across heterogeneous sequence distributions often causes negative transfer. We introduce FEAST (FEderated Activation STeering), a framework that enables clients to collaboratively construct steering directions without exchanging raw sequences. To overcome data heterogeneity, we propose target-domain purification and dynamic layer-wise anchor aggregation. Experiments on toxicity and immunogenicity under domain shift and data asymmetry show that our approach effectively mitigates negative transfer and outperforms single-client local steering. Notably, the framework requires only a single communication round of compact vectors--exchanging no gradients or model parameters--providing a highly efficient, privacy-preserving solution for collaborative protein design.